acceptodds
Under review as a conference paper at ICLR 2027

HyperSync: Hyperbolic Feature Matching for Multi-Agent Spatial Synchronization

Abstract

Multi-agent spatial synchronization is a critical prerequisite for cooperative perception and information fusion in autonomous driving. Existing methods typically establish cross-agent feature correspondences using metrics such as inner products and bilinear similarities in euclidean feature spaces, or perform association using object trajectories and motion-consistency constraints. When substantial viewpoint differences, incomplete observations, or only a limited number of co-visible objects exist among different agents, existing methods have limited ability to distinguish true correspondences, false correspondences, and outliers, making them prone to erroneous associations and consequently degrading the stability of relative pose estimation and spatial alignment. To address these challenges, we propose HyperSync, a multi-agent spatial synchronization method based on hyperbolic-space representations. By exploiting the negative curvature and nonlinear distance properties of hyperbolic space, HyperSync maps candidate matching features extracted from different agents into the Poincaré ball model and constructs a cross-agent feature similarity metric based on hyperbolic distance, thereby enhancing the discriminability of feature correspondences under weak-overlap and sparse-co-visibility conditions. On this basis, spatial alignment among multi-agent observations is achieved through correspondence filtering and relative pose estimation. The proposed method is evaluated on the OPV2V, V2X-Sim and DAIR-V2X datasets. Experimental results demonstrate that, compared with existing methods, the proposed method achieves more accurate feature matching and relative pose estimation and exhibits better robustness under substantial viewpoint differences and conditions with only a limited number of co-visible objects.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Autonomous Driving、Point Cloud Matching、Computer Vision、Smart City

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_D5KDOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper tackles prior-free relative pose estimation for cooperative perception. The proposed framework maps pair-conditioned BEV descriptors into a Poincaré ball, establishes correspondences via hyperbolic geodesic distances, and refines them using sub-pixel interpolation, geometric consistency filtering, robust SE(2) estimation, and an inlier-ratio gate. Evaluations are conducted on OPV2V, V2X-Sim, and DAIR-V2X against baselines including CBM+SVD, VIPS, V2X-Reg++, and CPCoCo, accompanied by a difficulty-stratified study on DAIR-V2X.

Critical strengths

The problem is practically important for V2X systems, and the empirical gains on DAIR-V2X are notable.

Critical weaknesses

My primary technical concern is that the core contribution (hyperbolic embedding) appears largely degenerate as implemented. Because every tangent vector is L2-normalized and scaled by a shared scalar prior to the exponential map, all embeddings are constrained to a sphere of fixed radius inside the Poincaré ball. On this manifold, the hyperbolic distance is strictly a monotonic function of the angular distance between descriptors, reducing the geodesic metric to a fixed reshaping of cosine similarity inside the InfoNCE logits (As stated at Line 261-263 in the paper). While Section 3.3 briefly acknowledges that mutual nearest neighbors coincide with cosine matching, the paper continues to attribute performance gains to negative curvature and exponential volume growth. In practice, radial variation is eliminated by construction. Without an ablation comparing this metric against standard cosine or Euclidean similarities under a tuned temperature parameter, the claim that hyperbolic geometry drives the matching performance remains unsupported.

Furthermore, the ablation study indicates an attribution issue regarding the source of improvements. Without Robust Sync and sub-pixel refinement, the raw descriptor matching yields mean errors of 4.451 m and 8.863 degrees on DAIR-V2X, performing substantially worse than CPCoCo (1.325 m and 3.688 degrees). Adding Robust Sync alone reduces the error to 0.758 m and 1.101 degrees, whereas sub-pixel refinement provides marginal value. This strongly suggests that the headline gains stem primarily from the post-processing pipeline rather than the learned hyperbolic representations. Because consistency filtering and robust gating are task-agnostic, a fair comparison requires evaluating whether baselines like CPCoCo achieve similar performance when equipped with the exact same post-processing. Additionally, V2X-Reg++ is used as a baseline across all tables but lacks formal citation and implementation details in the main text.

Questions for authors

None

Reviewer_8ejfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

HyperSync estimates planar relative poses between LiDAR-equipped agents without localization priors at inference. It constructs pair-conditioned BEV descriptors, embeds normalized features in a Poincaré ball, and learns correspondences using hyperbolic-distance InfoNCE alongside detection supervision. Mutual matches undergo local soft-argmax refinement, geometric consistency filtering, robust SE(2) fitting, and pose rejection. Experiments on OPV2V, V2X-Sim, and DAIR-V2X report lower mean translation and rotation errors than the evaluated baselines, with the largest gains on DAIR-V2X. A difficulty-stratified evaluation also reports improved overall success when abstentions count as failures. The ablations examine robust synchronization and sub-pixel refinement.

Critical strengths

The coverage-aware DAIR-V2X evaluation is convincing. On the 503 Hard pairs, HyperSync gets 90.1% coverage and 86.5% Native success, vs 55.3% for CBM+SVD and 19.5% for CPCoCo. Since rejected pairs count as failures, this isn’t just an easier accepted subset.

Critical weaknesses

The hyperbolic contribution isn’t isolated. Normalization puts embeddings at a shared radius, so mutual-NN ranking matches cosine for fixed descriptors. Hyperbolic distance may still help through loss shaping/refinement, but there’s no retrained cosine/Euclidean baseline with the same architecture and pipeline. Table 3 also keeps hyperbolic matching throughout, so it’s unclear what the metric adds beyond refinement and robust estimation.

The biggest ablation gain is also confounded by rejection. Robust Sync improves RTE/RRE a lot, but coverage/Native success are omitted, so I can’t tell whether pose recovery improved or hard pairs were just withheld. Eq. (38) also needs clarification: for SO(2), the Frobenius norm is sqrt(2) times yaw error, and it’s unclear if this affects the reported rotation errors and thresholds.

Questions for authors
  1. How does hyperbolic training compare with retrained cosine/Euclidean InfoNCE, with tuned temperatures and everything else fixed?
  2. Can Table 3 report coverage and Native success for every variant, and separate correspondence filtering from whole-pair rejection?
  3. Is Eq. (38) implemented literally, or is absolute wrapped yaw error used?
Reviewer_NiRDOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

HyperSync estimates pairwise planar relative poses for cooperative LiDAR perception without localization priors. It constructs pair-conditioned BEV descriptors and trains their correspondence matching using hyperbolic-distance-based contrastive learning. Discrete matches are refined to continuous coordinates, followed by geometric consistency filtering, robust SE(2) estimation, and rejection of insufficiently supported poses. The claimed contributions are hyperbolic matching, sub-pixel refinement, and robust synchronization. Validation covers OPV2V, V2X-Sim, and DAIR-V2X, with pose-error metrics, difficulty-stratified evaluation, and ablations of the two post-processing components. The authors report improvements over object-level and feature-level baselines.

Critical strengths

None

Critical weaknesses
  1. All embeddings share a fixed radius, making hyperbolic distance a monotonic transformation of cosine similarity and preserving nearest-neighbor rankings, as the authors acknowledge. Any benefit must therefore be demonstrated through training or score calibration. However, the ablations evaluate only post-processing, with the largest improvements arising from Robust Sync. Without controlled cosine/Euclidean baselines using the same architecture, tuned temperatures, and post-processing, the results do not establish that the paper’s central hyperbolic component provides a meaningful advantage.

  2. Equation (38) yields (\sqrt{2}) times the usual yaw-angle error, leaving the implementation and 3° success criterion unclear. Table 2 also contains values inconsistent with the stated identity (Native=Cover \times CSR): for Easy CPCoCo, the reported percentages imply approximately 33.7%, whereas Native is listed as 32.8%. Furthermore, accepted-only pose statistics lack coverage information in the main comparison and ablation tables. These issues affect both metric correctness and whether improvements reflect better estimation or selective rejection.

  3. I can only see a full screen of green in the supplementary material video.

Questions for authors
  1. Clarification: Which RRE formula is implemented for HyperSync and each baseline? Does the 3° success threshold refer to the actual yaw difference or the unnormalized matrix-logarithm norm?

  2. Clarification: Can the authors provide the candidate, valid-output, and successful-output counts underlying Table 2 and explain the inconsistencies between Cover, CSR, and Native? Were these quantities aggregated across multiple runs?

  3. Assessment-critical evidence: Does hyperbolic training outperform independently temperature-tuned cosine/Euclidean training when the backbone, projection, training budget, refinement, and rejection procedure are held fixed? Please provide coverage and full-set success alongside pose errors, including for the post-processing ablations.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.